Hacker News Top 10
English Edition

1. Facebook now charging to post more than 2 links a month (112 points by AshleysBrain)

   The linked Mastodon post claims that Facebook now charges users to post more than two links per month. The preview is mostly a truncated complaint and does not show Facebook’s policy text, pricing, or supporting evidence. The post’s central claim is a new restriction or fee on link sharing. The supplied material does not establish the exact scope or whether the claim is accurate.

2. DeepSeek just open-sourced its inference engine (1579 points by ggcr)

   The article announces that DeepSeek has open-sourced DeepInfer, the inference framework used for its large-scale production services. An inference engine is the software that runs a trained model and produces answers to requests. The repository says it provides offline inference with core features needed for serving at scale, supports DeepSeek-V4.1-Flash, and includes optimized kernels, model and sampler runtime, and parallelism features. It also includes separate requirements files for NVIDIA and Ascend hardware, plus tests and examples. The preview does not provide benchmarks, license details, or independent performance data.

3. EMudhra Intermediate CA Compromise (129 points by galadran)

   The article reports that private keys for eight eMudhra intermediate certificate authority certificates were compromised. An intermediate certificate authority signs certificates under a trusted root, so compromise can affect many downstream systems. The reporter signed a message with each compromised key and reported the issue to root program teams on October 17. Firefox and Chrome blocked the certificates on October 17, Apple on October 19, and Microsoft on October 20, while downstream trust stores began updates. An attacker with those keys could impersonate websites or services.

4. Apple broke an essential presentation feature of iPads (225 points by evanb)

   The author describes a regression in iPadOS 27 that breaks a presentation workflow used for teaching. The setup uses AirPlay to send a handout or presentation to an external screen while the iPad is also used for notes, polls, and controlling a PowerPoint deck. With multiple windows open, tapping outside the presenting app now makes the external display show the whole iPad screen. This exposes private notes or poll results to the audience and disrupts the presentation.

5. Ofcom Concedes Defeat in the 4chan Case (76 points by iamnothere)

   The article argues that Ofcom has effectively conceded defeat in its enforcement action involving 4chan. It quotes Ofcom’s chief executive saying that if foreign courts interfere, there is a point where Ofcom may not be able to recover money. The article says a U.S. court’s preliminary view in 4chan v. Ofcom suggests that enforcing Ofcom’s fines in the United States would face a cold reception. The piece is a legal and political commentary rather than a full court ruling.

6. How I got over 300 Tor exit nodes removed (88 points by kabakabakaba)

   The article describes a coordinated attack in which 333 infected Tor exit nodes, roughly 10% of all exit relays, intercepted SSH connections and collected unencrypted mail traffic. The attackers had access since August and deployed an OpenSSH backdoor and credential stealer called Ebury 1.8.3. They blocked some ports to focus on unencrypted mail and redirected SSH connections on port 22 to interception servers. The author encountered the SSH redirection, saw a host-key warning, and worked to get more than 300 exit nodes removed.

7. Triplet of 'ghostly' galaxies hints at what dark matter is and isn't (24 points by gumby)

   The article discusses three faint galaxies, DF2, DF9, and DF4, that appear to lack dark matter. Dark matter is the invisible material inferred from its gravitational effects on visible matter and large structures. The article says an ancient collision may have stripped these galaxies of dark matter. This observation is presented as trouble for alternative gravity theories, because those theories struggle to explain galaxies that behave as if dark matter is missing. The result supports the idea that dark matter is a separate substance, though the explanation remains a hypothesis.

8. My Friend Bob Cringely (59 points by zinxq)

   The article is a personal recollection by Paul Tyma about his friend Bob Cringely. It opens in 1996, when Cringely called Tyma after reading several Java books and praised Tyma’s book as the best. Tyma was skeptical because so few Java books existed at the time, but he recognized that Cringely knew how to appeal to a young programmer’s desire for recognition. The supplied preview only covers the opening of the story.

9. Launch HN: Respan (YC W24) – Open-source Rust AI gateway built around plugins (13 points by Andydy25)

   The Hacker News Launch entry introduces Respan, described as an open-source Rust AI gateway built around plugins. A gateway is a service that sits between an application and one or more model providers. The supplied content contains no description of Respan’s features, license, performance, or maturity. Because no preview text was available, its claims cannot be verified from the supplied material.

10. 2026 is an exciting year for the Rust compiler (and me) (38 points by panstromek)

    The post argues that 2026 is an unusually important year for the Rust compiler because several long-running efforts are reaching completion. The largest example is a next-generation trait solver, a reimplementation of a major part of Rust’s type system that has been in development for almost ten years. The author says it fixes soundness bugs and unblocks many language features and fixes, including work in async Rust. The preview frames 2026 as a payoff year for long-term compiler work rather than a sudden increase in effort.

AI/ML Insights & Trends

1. Open-source inference software is becoming a first-class AI deployment option

   • Trend/insight: DeepSeek open-sourced DeepInfer, the inference framework used for its production services, including offline inference, optimized kernels, runtime code, and tests (Article 2). This shows a model provider publishing not only model weights but also the software needed to serve them.
   • Why it matters for AI development: Teams building AI applications often depend on hosted APIs. An open inference stack gives another option: self-host a model and control the serving environment.
   • Implications and practical tips: Before replacing a hosted API, check the license, supported hardware, maintenance activity, and benchmark results. The supplied preview does not include performance numbers, so test the engine on your own workload.

2. Serving large models is a systems-engineering problem

   • Trend/insight: DeepInfer lists parallelism modes for different model layers, including mixture-of-experts layers, plus expert load balancing (Article 2). A mixture-of-experts model contains many specialized sub-networks and sends each input to only a few of them, so serving must manage routing and load.
   • Why it matters for AI development: Latency and cost depend on how work is split across accelerators and how evenly requests use different model parts. A local demo can hide these production issues.
   • Implications and practical tips: If self-hosting, test under realistic concurrency and monitor accelerator utilization and queueing. Start with a smaller model or a managed service if you cannot operate this infrastructure.

3. Hardware portability is a real integration cost

   • Trend/insight: The DeepInfer repository includes separate requirements files for NVIDIA and Ascend hardware (Article 2). This suggests the inference stack is intended for more than one accelerator family, at least at the dependency level. The preview does not show equal support or performance across vendors.
   • Why it matters for AI development: AI deployment choices are tied to available chips, drivers, and libraries. Moving between hardware vendors can require setup work and retesting.
   • Implications and practical tips: Treat hardware support as both a procurement and engineering question. Verify supported versions, container images, and benchmark parity before committing to a deployment target.

4. AI gateways are emerging as a control layer, but evidence is thin

   • Trend/insight: The Launch HN title describes Respan as an open-source Rust AI gateway built around plugins (Article 9). A gateway sits between your application and model providers and can potentially centralize routing, logging, or policy. The supplied content has no details, so this is an observation from a title, not proof of a mature trend.
   • Why it matters for AI development: As applications call multiple models or APIs, a central layer can reduce duplicated integration code and make controls easier to enforce.
   • Implications and practical tips: If considering a gateway, check the plugin API, supported providers, authentication, logging, latency, and license. Do not assume the title’s claims are backed by the missing content.

5. AI services still depend on ordinary network and certificate security

   • Trend/insight: Article 3 reports compromised intermediate certificate keys that could let an attacker impersonate websites or services. Article 6 reports compromised Tor exit nodes intercepting SSH connections and collecting unencrypted mail. These incidents are not AI-specific, but AI applications commonly call hosted APIs and expose web services over the same trust infrastructure.
   • Why it matters for AI development: A model endpoint is only as trustworthy as the TLS connection, certificate chain, and server identity protecting it. Weak network security can expose prompts, data, or credentials regardless of the model.
   • Implications and practical tips: Keep trust stores and dependencies updated, avoid unencrypted protocols, monitor host keys and certificates, and limit what sensitive data is sent to external services. When self-hosting, apply the same network controls you would use for any production service.

Analysis generated by deepseek-flash